Breast cancer risk prediction method and system fusing exposure parameters and image time sequence characteristics

Abstract

The invention provides a breast cancer risk prediction method and system fusing exposure parameters and image time sequence characteristics, and the method comprises the steps: collecting mammary gland molybdenum target images of a same to-be-predicted breast cancer patient at different time points, and carrying out the spatial registration and standardization processing of the mammary gland molybdenum target images, and obtaining a standardized image; based on a U-Net model, segmenting the standardized image to obtain a breast parenchyma region image; based on the breast parenchyma region image, adopting an OfficientNet-B4 network model to obtain a global depth feature vector; and obtaining a two-dimensional exposure parameter difference vector based on imaging parameters during mammary gland molybdenum target image acquisition, splicing the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence, inputting the two-dimensional sequence into the LSTM network model, obtaining an onset risk probability value of a to-be-predicted breast cancer patient, and completing prediction of the breast cancer risk. According to the method, the risk identification capability in longitudinal mammary image analysis can be remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image analysis, and specifically relates to a breast cancer risk prediction method and system that integrates exposure parameters and image timing characteristics. Background Art

[0002] Breast cancer is one of the most common malignant tumors among women worldwide. Early detection and accurate diagnosis are crucial for improving cure rates and reducing mortality. Mammography, the most widely used breast disease screening tool, can effectively detect early-stage lesions such as microcalcifications and masses.

[0003] In practice, doctors typically compare breast images of the same patient at different time points to determine whether suspicious changes in breast tissue structure have occurred and assess potential disease risks. However, the complex structure of the breast, the large individual differences, and the diverse morphology of lesions, coupled with differences in imaging conditions (such as compression thickness and radiation dose) at different examination time points, mean that image changes may be due to both true pathological evolution and "pseudo-changes" caused by changes in imaging parameters. This interference can challenge doctors' judgments or the predictive accuracy of models, especially in automated risk prediction systems based on image sequences.

[0004] In recent years, deep learning technology has made significant progress in medical image analysis, particularly in breast image classification, detection, and segmentation. However, most existing methods focus on static analysis of single-phase images, lacking the ability to model evolving trends across multiple time points. They also generally ignore the impact of imaging parameter changes on image performance. This results in insufficient model robustness in longitudinal imaging analysis tasks, making it difficult to effectively support high-precision breast cancer risk prediction.

[0005] Therefore, there is an urgent need for an intelligent prediction method that can simultaneously consider the temporal evolution characteristics of images and changes in imaging parameters, improve the model's ability to identify real pathological changes, and accurately assess the development trend and risk level of breast cancer, thereby providing more scientific and intelligent decision-making support for screening, individualized intervention, and long-term follow-up management. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention provides a breast cancer risk prediction method and system that integrates exposure parameters and image timing characteristics, aiming to solve the key problem in existing breast image analysis methods that it is difficult to accurately distinguish between "pseudo-changes caused by changes in imaging parameters" and "image changes caused by real pathological evolution."

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A breast cancer risk prediction method integrating exposure parameters and image temporal characteristics, the method comprising:

[0009] S1, collecting mammographic images of the same patient with breast cancer to be predicted at different time points, performing spatial registration and standardization processing on the mammographic images to obtain standardized images;

[0010] S2. Segmenting the standardized image based on a U-Net model to obtain a breast parenchymal area image;

[0011] S3. Based on the breast parenchyma area image, using the EfficientNet-B4 network model, extract global depth features to obtain a global depth feature vector;

[0012] S4. Obtaining a two-dimensional exposure parameter difference vector based on the imaging parameters during the mammographic image acquisition;

[0013] S5. Splicing the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence;

[0014] S6. Based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain a probability value of the breast cancer risk of the patient to be predicted, thereby completing the prediction of the breast cancer risk.

[0015] Preferably, the spatial registration and standardization processing of the mammographic target image in S1 includes:

[0016] S11, performing spatial registration on the mammography image: using affine transformation, including rotation, translation, and scaling, to obtain a registered mammography image;

[0017] S12. Standardize the registered mammography images: grayscale normalization, linearly scaling pixel values ​​to the interval [0, 1]; contrast enhancement, using contrast-limited adaptive histogram equalization with a window size of 8 × 8 and a clipping limit of 2.0; noise removal, using a 3 × 3 window median filter.

[0018] Preferably, in S2, the standardized image is segmented based on the U-Net model to obtain the breast parenchymal area image, including: using a pre-trained U-Net model to segment the standardized image, removing the pectoral muscle, background and label information, and obtaining the breast parenchymal area image, and the segmented image is cropped and uniformly scaled to 380×380 pixels.

[0019] Preferably, the pre-trained U-Net model includes:

[0020] The pre-trained U-Net model is initialized using the U-Net network weights pre-trained on the breast medical image segmentation dataset INBreast;

[0021] Freeze the first two convolutional layers of the encoder of the pre-trained U-Net model and fine-tune the decoder and high-level semantic modules;

[0022] Introducing the joint optimization strategy of Dice loss and binary cross entropy;

[0023] The training was performed using a manually annotated breast ROI training set combined with data augmentation methods such as horizontal / vertical flipping, random rotation, and Gaussian perturbation.

[0024] Preferably, in S3, based on the breast parenchyma area image, an EfficientNet-B4 network model is used to extract global depth features and obtain a global depth feature vector, including:

[0025] S31, deleting the top classifier in the EfficientNet-B4 network model, freezing the first 50% of the convolutional layers of the EfficientNet-B4 network model, and fine-tuning the last 50% of the convolutional layers to obtain an improved EfficientNet-B4 network model;

[0026] S32. Input the breast parenchymal area image into the improved EfficientNet-B4 network model, extract the global depth features of the breast parenchymal area image through the GAP layer, and obtain the global depth feature vector.

[0027] Preferably, in S31, the top classifier in the EfficientNet-B4 network model is deleted, the first 50% of the convolutional layers of the EfficientNet-B4 network model are frozen, and the last 50% of the convolutional layers are fine-tuned to obtain an improved EfficientNet-B4 network model, including:

[0028] S311, model structure adjustment: delete the top classifier in the EfficientNet-B4 network model, and add a new global average pooling layer and a fully connected layer;

[0029] S312, parameter freezing and unfreezing: Freeze the first 50% of the convolutional layers of the EfficientNet-B4 network model, set model.layers[:j / 2].trainable = False, where j is the total number of convolutional layers of the EfficientNet-B4; unfreeze the last 50% of the convolutional layers of the EfficientNet-B4 network model, allow the last 50% of the convolutional layers to update their weights during training, and set model.layers[j / 2:].trainable = True;

[0030] S313. Fine-tune the training strategy: set the learning rate, use the AdamW optimizer, apply rotation, scaling, and brightness adjustment enhancement to the breast parenchymal area images, and use binary cross entropy as the loss function.

[0031] Preferably, the step of obtaining a two-dimensional exposure parameter difference vector based on imaging parameters during mammography image acquisition in S4 includes:

[0032] S41, obtaining imaging parameters compression thickness and radiation dose during two acquisitions of the mammographic target images;

[0033] S42, respectively calculating the difference vectors of the compression thickness and the radiation dose during two acquisitions to obtain a compression thickness difference vector and a radiation dose difference vector;

[0034] S43 , performing Min-Max normalization processing on the compression thickness difference vector and the radiation dose difference vector to obtain the two-dimensional exposure parameter difference vector.

[0035] Preferably, the step of splicing the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence in S5 includes:

[0036] S51, performing normalization processing on the two-dimensional exposure parameter difference vector;

[0037] S52, concatenating the normalized two-dimensional exposure parameter difference vector with the global depth feature vector at the corresponding time point to obtain a joint feature vector;

[0038] S53: Combine the joint feature vectors in time sequence to obtain the two-dimensional sequence.

[0039] Preferably, in S6, based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain the disease risk probability value of the breast cancer patient to be predicted, and complete the prediction of the breast cancer risk, including:

[0040] S61, introducing an attention mechanism into the bidirectional LSTM network model;

[0041] S62. Calculate and obtain the importance weight of the two-dimensional sequence through the attention mechanism;

[0042] S63: Weighting and mapping the importance weights, and obtaining the risk probability value of the breast cancer patient to be predicted through a Sigmoid activation function, thereby completing the prediction of the breast cancer risk.

[0043] The present invention also provides a breast cancer risk prediction system that integrates exposure parameters and image temporal features, the system comprising: an image preprocessing module, an image segmentation module, an image feature extraction module, an exposure parameter feature construction module, a feature fusion module, and a prediction module;

[0044] The image preprocessing module is used to collect mammographic images of the same patient with breast cancer to be predicted at different time points, perform spatial registration and standardization processing on the mammographic images, and obtain standardized images;

[0045] The image segmentation module is used to segment the standardized image based on the U-Net model to obtain a breast parenchymal area image;

[0046] The image feature extraction module is used to extract global depth features based on the breast parenchyma area image and adopt the EfficientNet-B4 network model to obtain a global depth feature vector;

[0047] The exposure parameter feature construction module is used to obtain a two-dimensional exposure parameter difference vector based on the imaging parameters during the mammographic image acquisition;

[0048] The feature fusion module is used to splice the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence;

[0049] The prediction module is used to obtain the morbidity probability value of the breast cancer patient to be predicted based on the two-dimensional sequence using a bidirectional LSTM network model to complete the prediction of breast cancer risk.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention provides a breast cancer risk prediction method and system that integrates exposure parameters and image timing characteristics. This method can significantly improve the risk identification capability in longitudinal breast image analysis and effectively avoid the risk of misjudgment caused by imaging differences. It has good practicality and promotion value in scenarios such as assisting individualized breast cancer prediction, intelligent screening systems, and long-term follow-up warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 Schematic diagram of the process of predicting breast cancer risk by integrating exposure parameters and image temporal characteristics according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Example 1

[0057] like Figure 1 As shown, the present invention provides a breast cancer risk prediction method that integrates exposure parameters and image temporal characteristics, including:

[0058] S1. Collect mammographic images of the same patient with breast cancer to be predicted at different time points, perform spatial registration and standardization on the mammographic images, and obtain standardized images;

[0059] S2. Based on the U-Net model, the standardized image is segmented to obtain the breast parenchymal area image;

[0060] S3. Based on the breast parenchymal area image, the EfficientNet-B4 network model is used to extract global depth features and obtain the global depth feature vector;

[0061] S4. obtaining a two-dimensional exposure parameter difference vector based on imaging parameters during mammography imaging acquisition;

[0062] S5, concatenating the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence;

[0063] S6. Based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain the risk probability value of breast cancer patients to be predicted, and complete the prediction of breast cancer risk.

[0064] Furthermore, the specific implementation process of the present invention is as follows:

[0065] S1. Collect mammographic images of the same patient with breast cancer to be diagnosed at different time points, perform spatial registration and standardization on the mammographic images, and obtain standardized images, including:

[0066] S11. Collect mammographic images of the same patient with breast cancer to be diagnosed at different time points and perform spatial registration of the mammographic images to ensure alignment of anatomical structures: Affine transformations, including rotation (±5°), translation, and scaling (ratio range 0.9 to 1.1), are used to obtain registered mammographic images.

[0067] S12. Standardize the registered mammography images: grayscale normalization, linearly scaling pixel values ​​to the range [0, 1]; contrast enhancement, using contrast-limited adaptive histogram equalization (CLAHE), with a window size of 8 × 8 and a clipping limit of 2.0; and noise removal, using a 3 × 3 window median filter to reduce high-frequency noise interference.

[0068] S2. Based on the U-Net model, the standardized image is segmented to obtain the breast parenchymal region image (ROI). This includes: using the pre-trained U-Net model to segment the standardized image, removing the pectoral muscles, background, and label information, and retaining only the region of interest, namely the breast parenchymal region image (ROI). To improve the segmentation accuracy and generalization ability of the model on mammography images, the pre-trained U-Net model was further fine-tuned and optimized for breast imaging, including:

[0069] The model was initialized using the U-Net network weights pre-trained on the public breast medical image segmentation dataset INBreast;

[0070] Freeze the first two convolutional layers of the model encoder and fine-tune only the decoder and high-level semantic modules to better adapt to the texture characteristics and tissue structure differences of mammography images.

[0071] The Dice loss and Binary Cross Entropy (BCE) joint optimization strategy is introduced to effectively alleviate the foreground-background pixel imbalance problem;

[0072] The model is trained using a manually annotated breast ROI training set combined with data augmentation methods such as horizontal / vertical flipping, random rotation, and Gaussian perturbation to enhance the model's robustness to changes in image brightness, position, and occlusion.

[0073] The fine-tuned U-Net model performs better in the tasks of breast parenchyma boundary extraction and pectoral muscle removal. Experiments have shown that it can increase the Dice coefficient from 0.83 of the original model to 0.89, effectively improving the accuracy and stability of subsequent feature extraction.

[0074] The segmented images are cropped and uniformly scaled to 380 × 380 pixels and used as input for subsequent feature extraction.

[0075] S3. Based on the breast parenchyma region image, the EfficientNet-B4 network model is used to extract global depth features and obtain a global depth feature vector, including:

[0076] S31. Delete the top classifier in the EfficientNet-B4 network model (based on ImageNet pre-trained weights), freeze the first 50% of the convolutional layers of the EfficientNet-B4 network model, and fine-tune only the last 50% of the convolutional layers to adapt to the texture characteristics of breast parenchymal area images, thereby obtaining an improved EfficientNet-B4 network model.

[0077] The original EfficientNet-B4 network model contains a top-level classifier for ImageNet classification. In the transfer learning task of the present invention, this top-level classifier is removed to adapt to breast cancer feature extraction.

[0078] Furthermore, the specific process of obtaining the improved EfficientNet-B4 network model is as follows:

[0079] ①Model structure adjustment

[0080] Remove the top classifier of the original model (the part after the GlobalAveragePooling2D layer).

[0081] Add a new global average pooling (GAP) layer and a fully connected layer (if needed) to output a feature vector.

[0082] ② Parameter freezing and unfreezing

[0083] Freeze the first 50% of the convolutional layers and set model.layers[:j / 2].trainable = False, where j is the total number of convolutional layers in EfficientNet-B4.

[0084] Unfreeze the last 50% of the convolutional layers: allow the second half of the layers to update weights during training (model.layers[j / 2:].trainable=True), where j is the total number of convolutional layers in EfficientNet-B4.

[0085] ③ Fine-tune training strategy

[0086] Learning rate setting: Use a lower learning rate (1e-5) to avoid destroying the effectiveness of pre-trained weights.

[0087] Optimizer selection: AdamW is used with weight decay to prevent overfitting.

[0088] Data augmentation: Rotation (±10°), scaling (0.9-1.1 times), and brightness adjustment (±10%) are applied to breast parenchymal region images to improve generalization.

[0089] Loss function: Use binary cross entropy (BCE).

[0090] ④Verification and tuning

[0091] Monitor the validation set loss and feature discrimination, and adjust the fine-tuning rounds to 50 rounds.

[0092] The improved EfficientNet-B4 network model has the following advantages:

[0093] Preserving general feature extraction capabilities: The frozen shallow convolutional layers have already learned low-level features such as edges and textures, avoiding overfitting caused by training from scratch;

[0094] Adapting to breast imaging characteristics: Fine-tuning the deep network can capture the specific characteristics of the breast parenchyma area and improve the targetedness of feature expression;

[0095] Computational efficiency optimization: Freezing some parameters reduces the amount of computation during training and accelerates model convergence;

[0096] Improved data efficiency: When medical data is limited, transfer learning significantly reduces the demand for labeled data.

[0097] S32. Input the breast parenchymal area image into the improved EfficientNet-B4 network model, extract the global depth features of the breast parenchymal area image through the GAP layer, and obtain a 1792-dimensional global depth feature vector.

[0098] S4. Based on the imaging parameters during mammography imaging, a two-dimensional exposure parameter difference vector is obtained, including:

[0099] S41. Obtain imaging parameters of compression thickness ΔT (in mm) and radiation dose ΔD (in mGy) during two acquisitions of mammographic images.

[0100] S42, respectively calculating the difference vectors of the compression thickness and the radiation dose during the two acquisitions, and obtaining the compression thickness difference vector and the radiation dose difference vector as auxiliary parameters for the model input;

[0101] S43. Perform Min-Max normalization processing on the compression thickness difference vector and the radiation dose difference vector to obtain a two-dimensional exposure parameter difference vector e=[△T, △D].

[0102] S5. Splice the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence, including:

[0103] The global depth feature vector (1792 dimensions) is concatenated with the two-dimensional exposure parameter difference vector (2 dimensions) to form a joint feature vector (1794 dimensions). The data from the two acquisition time points form a time series of length 2, which is fed into the subsequent bidirectional LSTM network as input.

[0104] When implementing the splicing of the "2D exposure parameter difference vector" and the "global depth feature vector", there are several key technical difficulties:

[0105] ①Data dimension mismatch

[0106] The global depth feature vector is high-dimensional (1792 dimensions), while the exposure parameter difference vector is low-dimensional (only 2 dimensions). Direct concatenation may result in: the exposure parameter information being submerged in the high-dimensional features; and the model having difficulty effectively learning the interaction between exposure parameters and image features.

[0107] ②Data semantic heterogeneity

[0108] Image features reflect visual information such as tissue structure and texture, while exposure parameters reflect the physical conditions of imaging, not the image content itself. These two types of information are different and can easily lead to semantic conflict or information interference when fused.

[0109] ③ Time series structural consistency issues

[0110] The present invention uses images at two time points and corresponding exposure parameters to construct a time sequence, i.e., a two-dimensional sequence, of length 2. If the splicing method is inappropriate, the time sequence may be destroyed or information may be misplaced.

[0111] To this end, the specific process of the present invention for splicing the "two-dimensional exposure parameter difference vector" and the "global depth feature vector" to obtain a two-dimensional sequence is as follows:

[0112] ① Input preparation

[0113] The present invention uses images at two time points and corresponding exposure parameters to extract breast parenchymal region images l for each time point t∈{t1, t2} t , obtain the corresponding imaging parameters (compression thickness T t , radiation dose D t ), from image l t Extract the global depth feature vector F t(1792 dimensions), and calculate the exposure parameter difference vector at the current time point t1 relative to another time point t2:

[0114]

[0115] Exposure parameter difference vector e t Min-Max normalization is performed to make its value range consistent with the depth feature vector.

[0116]

[0117] Where x represents the original exposure parameter difference (compression thickness difference △T or radiation dose difference △D), x min with x max Respectively represent the minimum and maximum values ​​of the parameter dimension in the training set, x ' It is a normalized value that falls between [0,1].

[0118] ② Time-step splicing

[0119] For each time point t, the normalized exposure parameter difference vector e t The corresponding image feature vector F t Splicing:

[0120] X t =F t ⊕e t

[0121] Among them, ⊕ represents the vector splicing operation, and the new joint feature vector X is obtained. t , with a dimension of 1794.

[0122] ③Build time series input

[0123] Combine the joint feature vectors of two time points in chronological order to form a time series of length 2:

[0124]

[0125] That is, X is a two-dimensional sequence with a shape of (2, 1794), which represents the joint features of the image features and exposure parameters at two time points.

[0126] S6. Based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain the risk probability value of breast cancer patients to be predicted, and complete the prediction of breast cancer risk, including:

[0127] S61. A bidirectional LSTM network is used to model the evolution trend of images and exposure parameters in the time series. The number of LSTM hidden units is set to 256, the time step is 2, and the attention mechanism is introduced into the bidirectional LSTM network model.

[0128] S62. Calculate the importance weight of the two-dimensional sequence through the attention mechanism;

[0129] S63. Weight the importance weights, input the weighted feature representation into the fully connected layer, map it to the 128-dimensional hidden layer, and then use the Sigmoid activation function to output the probability value of the breast cancer risk of the patient to be predicted, indicating the predicted result of the breast cancer risk, thereby completing the prediction of the breast cancer risk.

[0130] The probability value can be converted into high / medium / low risk levels, or used for binary classification judgment (with / without obvious risk).

[0131] Among them, the training strategy of the bidirectional LSTM network model is as follows:

[0132] Loss function: Use Binary Cross-Entropy Loss.

[0133] Optimizer: AdamW, initial learning rate 1e-4, weight decay 1e-5.

[0134] Regularization: Dropout (ratio 0.5) is added to the fully connected layer.

[0135] Data partitioning: The training set, validation set, and test set were divided into a ratio of 7:2:1, and five-fold cross-validation was used to evaluate the model stability.

[0136] Training rounds: Set the initial number of training rounds to 100, and dynamically adjust the learning rate based on the performance of the validation set in each round.

[0137] Evaluation indicators include: AUC, accuracy, sensitivity, specificity and F1-score.

[0138] Comparative experiment

[0139] Baseline model: LSTM model using only image features (without exposure parameters).

[0140] Ablation experiment: Remove the exposure parameter and attention mechanism respectively to analyze the contribution of each module to performance.

[0141] Specifically, the comparison model design is shown in Table 1:

[0142] Table 1

[0143]

[0144] The experimental results are shown in Table 2:

[0145] Table 2

[0146]

[0147] From the above experimental results, we can see that:

[0148] ①Introducing exposure parameters significantly improved model performance

[0149] Compared with M1, M2 has significant improvements in all indicators, with AUC increased by 3.8% and F1-score increased by 4.9%. This shows that fusing the exposure parameter difference vector with image features helps the model better distinguish "true lesions" from "false changes" and improves discrimination accuracy.

[0150] ②Introducing the attention mechanism further optimizes the model performance

[0151] Compared with M2, M3 has a 2.4% improvement in AUC and a 2.4% improvement in F1-score; the attention mechanism enables the model to dynamically focus on more discriminative change trends between two time points, improving the interpretability and robustness of the model.

[0152] ③The use of bidirectional LSTM significantly enhances the time series modeling capability

[0153] Compared with M3, M4's AUC increased by 1.7% and F1-score increased by 1.9%; the bidirectional structure effectively captures the interdependence between previous and subsequent time points, improving the model's ability to understand the pathological evolution process.

[0154] In summary, the breast cancer risk prediction method provided by the present invention, which integrates exposure parameters and image timing characteristics, can significantly improve the risk identification ability in longitudinal breast image analysis, effectively avoid the risk of misjudgment caused by imaging differences, and has good practicality and promotion value in scenarios such as assisting individualized breast cancer prediction, intelligent screening systems, and long-term follow-up warnings.

[0155] Example 2

[0156] The present invention also provides a breast cancer risk prediction system that integrates exposure parameters and image temporal features. The system includes: an image preprocessing module, an image segmentation module, an image feature extraction module, an exposure parameter feature construction module, a feature fusion module, and a prediction module.

[0157] An image preprocessing module is used to collect mammographic images of the same patient with breast cancer to be predicted at different time points, perform spatial registration and standardization on the mammographic images, and obtain standardized images;

[0158] An image segmentation module is used to segment the standardized image based on the U-Net model to obtain an image of the breast parenchyma area;

[0159] The image feature extraction module is used to extract global depth features based on breast parenchymal area images and obtain global depth feature vectors using the EfficientNet-B4 network model;

[0160] An exposure parameter feature construction module is used to obtain a two-dimensional exposure parameter difference vector based on the imaging parameters during mammography imaging acquisition;

[0161] The feature fusion module is used to splice the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence;

[0162] The prediction module is used to obtain the risk probability value of breast cancer patients to be predicted based on the two-dimensional sequence and use the bidirectional LSTM network model to complete the prediction of breast cancer risk.

[0163] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A breast cancer risk prediction method integrating exposure parameters and image temporal characteristics, characterized in that: The method comprises: S1, collecting mammographic images of the same patient with breast cancer to be predicted at different time points, performing spatial registration and standardization processing on the mammographic images to obtain standardized images; S2. Segmenting the standardized image based on a U-Net model to obtain a breast parenchymal area image; S3. Based on the breast parenchyma area image, using the EfficientNet-B4 network model, extract global depth features to obtain a global depth feature vector; S4. Obtaining a two-dimensional exposure parameter difference vector based on the imaging parameters during the mammographic image acquisition; S5. Splicing the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence; S6. Based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain a probability value of the breast cancer risk of the patient to be predicted, thereby completing the prediction of the breast cancer risk.

2. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, characterized in that: The step S1 performs spatial registration and standardization on the mammography image, including: S11, performing spatial registration on the mammography image: using affine transformation, including rotation, translation, and scaling, to obtain a registered mammography image; S12. Standardize the registered mammography images: grayscale normalization, linearly scaling pixel values ​​to the interval [0, 1]; contrast enhancement, using contrast-limited adaptive histogram equalization with a window size of 8 × 8 and a clipping limit of 2.0; noise removal, using a 3 × 3 window median filter.

3. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, characterized in that: In S2, the standardized image is segmented based on the U-Net model to obtain the breast parenchymal area image, including: using a pre-trained U-Net model to segment the standardized image, removing the pectoral muscle, background and label information, and obtaining the breast parenchymal area image, and the segmented image is cropped and uniformly scaled to 380×380 pixels.

4. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 3, wherein: The pre-trained U-Net model includes: The pre-trained U-Net model is initialized using the U-Net network weights pre-trained on the breast medical image segmentation dataset INBreast; Freeze the first two convolutional layers of the encoder of the pre-trained U-Net model and fine-tune the decoder and high-level semantic modules; Introducing the joint optimization strategy of Dice loss and binary cross entropy; The training was performed using a manually annotated breast ROI training set combined with data augmentation methods such as horizontal / vertical flipping, random rotation, and Gaussian perturbation.

5. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, characterized in that: In S3, based on the breast parenchyma area image, the EfficientNet-B4 network model is used to extract global depth features and obtain a global depth feature vector, including: S31, deleting the top classifier in the EfficientNet-B4 network model, freezing the first 50% of the convolutional layers of the EfficientNet-B4 network model, and fine-tuning the last 50% of the convolutional layers to obtain an improved EfficientNet-B4 network model; S32. Input the breast parenchymal area image into the improved EfficientNet-B4 network model, extract the global depth features of the breast parenchymal area image through the GAP layer, and obtain the global depth feature vector.

6. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 5, characterized in that: In S31, the top classifier in the EfficientNet-B4 network model is deleted, the first 50% of the convolutional layers of the EfficientNet-B4 network model are frozen, and the last 50% of the convolutional layers are fine-tuned to obtain an improved EfficientNet-B4 network model, including: S311, model structure adjustment: delete the top classifier in the EfficientNet-B4 network model, and add a new global average pooling layer and a fully connected layer; S312, parameter freezing and unfreezing: Freeze the first 50% of the convolutional layers of the EfficientNet-B4 network model, set model.layers[:j / 2].trainable = False, where j is the total number of convolutional layers of the EfficientNet-B4; unfreeze the last 50% of the convolutional layers of the EfficientNet-B4 network model, allow the last 50% of the convolutional layers to update their weights during training, and set model.layers[j / 2:].trainable = True; S313. Fine-tune the training strategy: set the learning rate, use the AdamW optimizer, apply rotation, scaling, and brightness adjustment enhancement to the breast parenchymal area images, and use binary cross entropy as the loss function.

7. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, characterized in that: The step S4 of obtaining a two-dimensional exposure parameter difference vector based on the imaging parameters during the mammographic image acquisition includes: S41, obtaining imaging parameters compression thickness and radiation dose during two acquisitions of the mammographic target images; S42, respectively calculating the difference vectors of the compression thickness and the radiation dose during two acquisitions to obtain a compression thickness difference vector and a radiation dose difference vector; S43 , performing Min-Max normalization processing on the compression thickness difference vector and the radiation dose difference vector to obtain the two-dimensional exposure parameter difference vector.

8. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, wherein: The step S5 of splicing the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence includes: S51, performing normalization processing on the two-dimensional exposure parameter difference vector; S52, concatenating the normalized two-dimensional exposure parameter difference vector with the global depth feature vector at the corresponding time point to obtain a joint feature vector; S53: Combine the joint feature vectors in time sequence to obtain the two-dimensional sequence.

9. The breast cancer risk prediction method integrating exposure parameters and image temporal characteristics according to claim 1, characterized in that: In S6, based on the two-dimensional sequence, a bidirectional LSTM network model is used to obtain the risk probability value of the breast cancer patient to be predicted, thereby completing the prediction of the breast cancer risk, including: S61, introducing an attention mechanism into the bidirectional LSTM network model; S62. Calculate and obtain the importance weight of the two-dimensional sequence through the attention mechanism; S63: Weighting and mapping the importance weights, and obtaining the risk probability value of the breast cancer patient to be predicted through a Sigmoid activation function, thereby completing the prediction of the breast cancer risk.

10. A breast cancer risk prediction system integrating exposure parameters and image temporal characteristics, the system comprising: Image preprocessing module, image segmentation module, image feature extraction module, exposure parameter feature construction module, feature fusion module and prediction module; The image preprocessing module is used to collect mammographic images of the same patient with breast cancer to be predicted at different time points, perform spatial registration and standardization processing on the mammographic images, and obtain standardized images; The image segmentation module is used to segment the standardized image based on the U-Net model to obtain a breast parenchymal area image; The image feature extraction module is used to extract global depth features based on the breast parenchyma area image and adopt the EfficientNet-B4 network model to obtain a global depth feature vector; The exposure parameter feature construction module is used to obtain a two-dimensional exposure parameter difference vector based on the imaging parameters during the mammographic image acquisition; The feature fusion module is used to splice the two-dimensional exposure parameter difference vector into the global depth feature vector to obtain a two-dimensional sequence; The prediction module is used to obtain the morbidity probability value of the breast cancer patient to be predicted based on the two-dimensional sequence using a bidirectional LSTM network model to complete the prediction of breast cancer risk.

Images